Surface Science Reports63(2008)487–513Contents lists available at ScienceDirectSurface Science Reports journal homepage:/locate/surfrepHigh-throughput heterogeneous catalysisDavid Farrusseng∗UniversitéLyon1,CNRS,UMR5256,IRCELYON,Institut de recherches sur la catalyse et l’environnement de Lyon,2avenue Albert Einstein,F-69626Villeurbanne,Francea r t i c l e i n f o Article history:Accepted16September2008 editor:W.H.Weinberg Keywords:CatalysisCombinatorial chemistryHigh-throughputKinetic modelingQSAR a b s t r a c tThis comprehensive review of the literature(over250references)deals with high-throughput experimentation in heterogeneous catalysis.Approaches to library design for catalyst discovery and optimization are described and discussed.Special focus is placed on advanced methods for knowledge discovery such as high-throughput kinetic modeling and QSAR.An inventory of successful case studies in catalysis is reported.Finally,recent developments in relevant electronic data and knowledge management are described.©2008Published by Elsevier B.V.Contents1.Introduction (488)2.Overview (488)3.Approaches to HT library design (489)3.1.The split&pool method (489)3.2.The hierarchical approach (490)3.3.Design of experiments(DoE)methodology(See also Figs.4and5) (492)3.4.Evolutionary algorithms (493)3.5.Evolutionary optimization using data mining tools (495)3.6.Summary (496)4.HT kinetic modeling (496)4.1.Reasons for conducting HT kinetic modeling (496)4.2.Technologies for HT kinetic modeling (497)4.3.Methodologies of HT kinetic modeling (497)4.4.Spatially-and time-resolved methods (499)5.The QSAR approach to catalysis (500)5.1.Generalities (500)5.2.Homogeneous catalysis and the QSAR approach (500)5.2.1.The QSAR concept (500)5.2.2.In silico generation of a virtual catalyst library (501)5.2.3.Choice and calculation of descriptors (501)5.2.4.QSAR modeling (501)5.3.Heterogeneous catalysis and the QSAR approach (502)5.4.HT physicochemical characterization for property quantification (502)5.5.Catalyst profiling using HT kinetic modeling of model reactions (503)5.6.Virtual screening through computational chemistry (504)Abbreviations:AniML,Analytical Information Markup Language;ANN,Artificial Neural Networks;ANOVA,Analysis Of Variance;DFT,Density Functional Theory;DoE, Design of Experiments;ee,enantioselectivity;FPA,Focalplane;FTIR,Fourier Transform Infrared;GA,Genetic Algorithm;GDC,Guided Data Capture;HT,High-throughput; HTE,High-throughput Experimentation;IT,Information Technology;KFE,Kinetic Fitting Engine;LRIS,Laboratory Research Informatic System;PCA,Principal Component Analysis;QM,Quantum Mechanics;QSAR,Quantitative Structure–Activity Relationship;QSPR,Quantitative Structure–Property Relationship;SCR,Selective Catalytic Reduction;SMEs,Small and Medium Enterprises;TPD,Temperature-programmed Desorption;TOF,Turnover Frequency;TON,Turnover Number;WGS,Water-gas Shift; WHSV,Weight Hourly Space Velocity;XML,Extensible Markup Language;XRD,X-ray Diffraction;XRF,X-ray Fluorescence.∗Tel.:+33472445365;fax:+33472445399.E-mail address:david.farrusseng@ircelyon.univ-lyon1.fr.0167-5729/$–see front matter©2008Published by Elsevier B.V.doi:10.1016/j.surfrep.2008.09.001488 D.Farrusseng/Surface Science Reports63(2008)487–5136.Discovery of catalytic materials by HT (504)6.1.Electrocatalysts for fuel cells (504)6.2.Selective hydrocarbon oxidation (504)6.3.Hydrogen production and purification (505)6.4.Automotive and refinery applications (505)6.5.Other applications (505)6.6.HT experimentation for zeolite synthesis and discoveries (505)7.Electronic infrastructure (506)7.1.Why automate data treatment? (506)7.2.Electronic open architecture for tool integration (506)7.2.1.HTE AG electronic platform (506)7.2.2.The NIST vision (506)7.2.3.Academic laboratories (507)7.2.4.Workflow-based electronic infrastructure for streamline data processing and knowledge management (507)7.2.5.Data normalization and e-standards in chemistry (509)8.Conclusions (509)Acknowledgements (510)References (510)1.IntroductionOver80%of commercial chemical processes involve the use of catalysis,with products as varied as chemicals,oil products, fertilizers,plastics,drugs and pharmaceuticals being made through catalytic steps.Catalysis is probably the most important means of producing modern chemicals;Europe’s chemical industry,for example,accounts for e1.5trillion,or14%,of this continent’s e10.5 trillion GDP(Gross Domestic Product).The likelihood of innovation in this field decreases,however, as catalytic chemical processes become increasingly mature. When new active solids are developed empirically,by trial-and-error processes employed on a few selected samples,the whole procedure is highly speculative and leads to a very slow rate of discovery for the industry in question.This research strategy based on exhaustive studies and complete understanding is also very time-consuming.Therefore,new research strategies have to be developed in order to produce breakthroughs and revitalize the field of chemical research.The high-throughput(HT)approach is a pragmatic alternative. It relies on the fast and systematic screening of libraries of diverse samples.This methodology is not new,since its origins can be found at different periods of the last century.One of the most striking examples is the discovery of the first ammonia synthesis catalyst by Mittasch et al.at BASF in1909followed by a‘‘systematic investigation of the periodic table’’with about 20,000experiments[1].This approach also appealed to K.Ziegler, who in the1950s applied it to the discovery of polymerization catalysts.The pragmatic approach of that time aimed at exploring and covering the periodic table;no references to HT screening or sample libraries could yet be made.We can trace the modern HT approach back to the pioneering work of Hanak in the1970s. He prepared and applied what we now call composition-spread or gradient libraries for research and development purposes at the RCA company laboratories.His work led to the successful entry of several new products onto the market[2].His vision of the experimental approach brought materials screening into the modern age:‘‘...the present approach to the search for new materials suffers from a chronic ailment,that of handling one sample at a time in the processes of synthesis,analysis and testing of properties.It is an expensive and time-consuming approach,which prevents highly-trained personnel from taking full advantage of its talents and keeps the tempo of discovery of new materials at a low level’’[3].The combinatorial principles employed in drug development were first applied to materials research in the early1990s by physi-cists and materials scientists at the Lawrence Berkeley National Laboratory of UC Berkeley.In fact,combinatorial materials sci-ence was recognized as a bona fide discipline only a few years later,following this team’s famous search for superconductors us-ing a materials library[4–7].By1997,the recently-formed Symyx Technologies had documented the state of the art of combinato-rial chemistry by publishing a library of over25,000distinct com-pounds[8].From2000onward,HT technology has been developed for and applied to an ever-increasing variety of materials,including electronic and magnetic materials,polymer-based materials,opti-cal materials,biomaterials,paints,drug formulations,detergents, cosmetics and glues,with the number of related publications and patents exploding accordingly.Today,HT experimentation has matured and is almost regarded as commonplace,its use in the development of new materials sometimes being omitted from a publication’s title or abstract,or even from the publication rge chemical companies (such as BASF,BP,Bayer,Degussa,DOW,DuPont,Exxon,GE, and UOP LLC)now generally have their own HT tools or labs. Meanwhile,smaller companies specialized in HT experimentation (such as Avantium,Bosch Lab Systems,hte AG,Symyx Technologies and Torial)have been founded,often enjoying spectacular growth in the space of the last ten years.Specialized companies such as these have succeeded thanks to their development of cutting-edge technologies,including hardware and software as well as their tight integration,that provide an impressive degree of throughput and productivity(number of samples screened per day and further decision making).In such a context of technological sophistication and high productivity,most academic groups have found themselves unable to compete in the race that is materials screening.Instead,public research centers can play a major role in HT by conducting fundamental research on domains as varied as synthetic methods,analytical tools,parallel in situ characterization,data mining and decision making processes and, finally–the focus of this review–screening strategies and methods.This review deals mainly with HT experimentation for hetero-geneous catalysis and also briefly discusses homogeneous cataly-sis.For other disciplines of materials science,the reader can refer to various reviews[9–19],books[20,21]and special issues[22–30]. An excellent summary of the state of the art for materials science has recently been published elsewhere[31].2.OverviewSince HT screening is a methodological approach,this review is divided into sections describing particular HT screening strategies and their associated strengths,issues,solutions,and case studies.What are the general issues affecting HT experimentation?D.Farrusseng/Surface Science Reports63(2008)487–513489Thanks to the modern screening techniques used in HT methods,dozens or even hundreds of experiments involving many variables can be performed at once.The inevitable combinatorial explosion that results leads to two urgent,fundamental questions relating to experimental design:which experiments are the most relevant to carry out,and what is the most efficient screening strategy?Data analysis is the next issue to come up after the experimental design is chosen.It is hardly possible for humans to fully evaluate results and statistical trends emanating from data sets involving more than four variables and20experiments.The issue of decision making takes on particular importance in such a scenario.In order to understand a catalytic process,one must be aware of the most relevant variables and combinations of variables affecting it.Catalysis,and chemistry in general, are matters of synergy and typically involve highly non-linear behaviors[32],as in metal–ligand or metal–support interactions for homogeneous and heterogeneous catalysis,respectively.Once the exploratory data analysis reveals whether it is possible to highlight positive interactions,and whether one can identify and quantify trends between variables and catalytic performances,a decision must be made—what is the most relevant experimental set to perform next?Computational methods employing mathematics,statistics and artificial intelligence are required for scientists to deal adequately with these three key issues influencing HT experimentation: experimental design,data analysis,and decision making.For over15years,these issues have been addressed in the field of drug discovery,leading to the emergence of a brand new domain of science with its own specialized journals.Section5deals with the application of such screening strategies and their related tools to HT catalyst design,as well as recent developments in the application of quantitative structure–activity relationship(QSAR) to homogeneous and heterogeneous catalysis,with an emphasis on the differences between organic and inorganic compounds in terms of descriptors.Kinetic modeling,on the other hand,reflects the scientist’s insight into the chemical kinetics and,therefore,provides useful information about catalyst behavior.It relates feedstock and operating conditions to reaction rates and corresponding effluent composition,thereby quantifying catalytic performances.Recent kinetic models also contain so-called catalyst descriptors,which specifically account for catalyst properties such as the number of sites and the reactant chemisorption enthalpy.Section4addresses the concepts,recent advances and limitations of kinetic modeling in HT experimentation.An appropriate electronic infrastructure for HT screening is absolutely necessary in order to prevent bottlenecks.Manual entry and cutting-and-pasting of data are to be minimized in order to limit the impact of erroneous entries and slowed-down experiments.Section7addresses these issues,with examples describing and illustrating current technology.Publications,handbooks and other documents available over the Internet,whether free of charge or at a fee,have rendered accessible a great deal of data that could possibly generate knowledge and assist in the decision-making process.That said,the direct capture of data is usually prevented due to the heterogeneity of data,as well as to a lack of standards regarding not just data format but also document format(*.pdf or*.html).The emergence of new technologies,concerted worldwide organizations for electronic standards and new business models appears to be changing this situation.This review does not deal only with catalyst design but also with the optimization of the process conditions due to the strong interplay between catalyst,reactor design and experimental testing conditions.Sections 3.1–3.3address the issues of HT strategies and library design,while Section4explores the benefits of using kinetic and transient approaches for catalyst design.3.Approaches to HT library designThe objectives of a targeted study,such as the discovery of entirely new compounds by exploring large search spaces or the fine optimization of a known catalyst[33,34],have a strong impact on the selection of an appropriate screening strategy and the associated information technology tools.Equipment constraints,synthesis feasibility and screening performances are also important factors to panies or HT departments have developed tailor-made solutions and entire workflows.The issues of HT infrastructure and tool integration as described in[35] will not be discussed here.This section describes the various screening strategies developed in industry and academia,in order from the simplest to the most complex strategies and algorithms, and to some extent from the most massive to the most qualitative screening.3.1.The split&pool methodIn the context of a primary screening aiming to identify hits in a very large search space,the split&pool approach, derived from pharmaceutical methodology,is especially well-suited.It is generally employed in ambitious catalyst discovery programs or when little is known about the target reaction and the class of materials to be studied.Even with a limited number of variables,several hundred samples can easily be obtained thanks to the combinatorial explosion.High analytical speed and overall throughput usually take precedence over the quality or density of information obtained.The split&pool approach enables the generation of every possible combination in a search space.While the generation methodology is quite simple,the major issue,due to its complexity, is the recognition of a molecule in a mixture.For this purpose, many different techniques have been developed to tag molecules, but such techniques cannot be directly applied to inorganic solids. Detailed below are two different tagging strategies,developed by UOP LLC and hte AG,that do allow tagging in the context of inorganic solids.In both strategies,one finds a massive parallel arrangement of micro-reaction chambers containing individual beads,each bead representing one catalyst as a member of a library of solid catalysts.This provides the advantages of easy catalyst handling(unlike the case of powders),very small quantities of metal precursors(typically100µg per catalyst)and a synthesis protocol that can be scaled up.At UOP,catalysts are tagged by using spatially addressable arrays such as microtiter plates(i.e.,96-well plates)[36].Each well contains a single catalyst bead which is indexed by four coordinates,namely the well-plate identity,row and column numbers and split and synthesis step.The combinatorial synthesis consists of a multi-step metal salt impregnation with intermediate drying.Once the metal salt solution is adsorbed onto the beads and dried,rows of beads from a given plate are transferred using a row-sorter to a set of receiving well plates.The sorting algorithm can be set in such a way that the sequence of row-and column-shuffling steps monitors the compositional redundancy of the resulting split–pool library.This process makes it possible for classic,inexpensive laboratory equipment to perform the synthesis of compositionally diverse libraries.The hte AG company has developed fast parallel post-analysis, an alternative to2D addressable layout tagging[37–40].The‘‘split’’step of the split&pool synthesis consists of dividing a few thousand beads into a number of equal portions placed on porcelain dishes. The beads,which are roughly1mm in diameter and are typical catalytic supports,such as alumina or titania,are impregnated with different metallic salt solutions varying by concentration or by metal nature,and then calcined.The beads are then recombined490 D.Farrusseng /Surface Science Reports 63(2008)487–513Fig.1.Split and pool synthesis (a),micro-bead reactor (b)after [39].Table 1Binary composition at the first screening.For each binary system,8different compositions are synthesized using a linear gradientapproach.(‘‘pooled’’)together and are well-mixed via shaking.Typically,the whole process is repeated several times.For example,for five different metal precursors at four different concentrations,the beads are split five times into four different containers (Fig.1).The total number of possible combinations from a mathematical perspective is 45=1024.That said,because the synthesis protocol cannot guarantee that each bead will follow a different path (making the preparation of some identical catalysts possible),it is recommended to start with a greater number of beads,with a ratio of 1.2–1.5with respect to the total theoretical number of combinations.At hte AG,up to 625beads can be tested individually by fast sequential testing methods in a specially-designed array of micro-reactors.The ‘‘hits’’identified are then post-analyzed by micro-X-ray fluorescence (or micro-XRF).In only a few minutes,commercial equipment can quantify the composition of an array of 100samples.The knowledge of the presence or absence of elements and of the range of concentration makes it possible to trace the synthesis path history for the beads.3.2.The hierarchical approachSome erroneous measurements inevitably occur during the primary screening due to the extreme miniaturization and the great number of experiments.Inactive catalysts can appear as ‘‘hits’’(false positives)and active catalysts that should be selected are missed (false negatives).In primary screening,false negatives (which would be missed forever)are far more problematic than are false positives (which would be discarded anyway following the secondary screening).With this fact in mind,Symyx has developed screening strategies to reduce the likelihood of false negatives.Over a two-week period in Symyx laboratories,a hierarchical gradient approach was used with MoVNb patented catalysts for ethane partial oxidation to acetic acid [41].This process led to the discovery of new dopants and to the successful rediscovery of existing catalysts.The primary screening is targeted to identify the best ternary combinations of redox metals from V,Mo,Cr,Mn,Fe,Co,Ni,Cu,Ag,Re,Sn,Sb,Ti and Bi.At the outset,30redox binary systems were investigated,with eight distinct samples synthesizedfor each binary system using a gradient approach consisting of a linear evolution of the composition (Table 1).Of the 30redox binaries (8-point gradients),MoV is by far the most active redox binary,while other active binaries are CrV,MnV,MnCr,CeV,CoCr,CoV,VTi and MoTi.Next,an additional element X was investigated within the MoV system,with X =Mg ,Cr ,Li ,Nb ,Mn ,Co ,Cu ,Fe ,Ni ,Zn ,Zr ,Sb ,Ag ,In or Ce,plus three proprietary dopants.For each ternary,12–15distinct compositions were screened.It was found that Nb,Ni,Sb,and Ce,as well as the three proprietary dopants,result in higher acetic acid productivity.Finally,among the 18ternary MoVX systems,eight of the most promising were tested again with 28different compositions per ternary system.In agreement with the literature,the MoVNb system was identified as very active,while several other metal dopants also produced hits.In summary,as the screening proceeds,the number of com-ponents in the catalytic system increases,formulation complexity increases and higher degrees of interactions are sought in a step-wise manner.The gradient library design approach also permits the efficient management of false negatives and false positives.The testing of similar compositions greatly reduces the risk of missing a hit,while allowing for the easy identification of false positives which afterward are not subjected to further screening.Symyx has demonstrated the value of this screening strategy for various types of catalysis,allowing the determination of the best formulations already reported in the literature and the discovery of new sys-tems [42].For example,the efficiency of Ni–Co–Nb and Ni–Ta–Nb oxide catalysts has successfully been proven for the ethane ox-idative dehydrogenation process at low temperatures [43],while supported-CoCr mixed oxide catalyst systems are proposed for VOC removal [44].Academic laboratories have also applied the gradient strategy.In a primary screening for the partial oxidation of isobutane [45–48],nine elements (V,Fe,Mo,Cr,Ta,Nb,Mn,Sb and Bi)were selected for an initial library design due to the promising partial oxidation properties reported in the literature for the associated oxides.As an alternative to a hierarchical approach,a single library was composed of the nine single oxides (AO x ),D.Farrusseng/Surface Science Reports63(2008)487–513491Fig.2.Emissivity-corrected IR-thermographic images of catalyst library during methanation of CO2at200◦C from[59].of double mixed oxides(A a B b O x)and of ternary mixed oxides (A a B b C c O x).The double mixed oxides consisted of all possible pairs with molar composition(0:1),(0.25:0.75),(0.5:0.5),(0.75:0.25) and(1:0)starting from the nine elements.The ternary mixed oxides consisted of combinations of three elements,always with a(0.33:0.33:0.33)molar composition.Most of the best dehydrogenation catalysts turned out to be Mn and Cr mixed oxides.In accordance with these results,four focused ternary systems–MoVSbO x,MoVFeO x,MoVBiO x and VBiSbO x–were studied in more detail using a dense grinding of the search space. The three elements were allowed to vary between0and100mol% in steps of10mol%,which led to the preparation of66samples per ternary.Other ternaries were investigated elsewhere[49].According to the authors,such composition-spread libraries should allow the identification of any existing local maxima. The best-performing mixed oxides among MoVSbO x,MoVFeO x, MoVBiO x and VBiSbO x were scaled up and tested in a secondary screening.The result was that the optimized composition, Mo10V10Sb80O x,surpassed the best reference catalyst in the literature.Several screening devices have been developed for the mapping of composition-spread libraries.Infrared(IR)thermography is an appropriate choice for the identification of active components in primary screening because,in principle,thousands of samples can be analyzed at the same time[11,50–61].This system records the temperature change arising from the exothermicity or endothermicity of the reaction(Fig.2).In principle,theselectivity Fig.3.Cross-section of the assembled reactor(left)and magnification of reaction chambers(right)from[70].Fig.4.Experimental planning for catalyst design(left).Headers correspond to the element names on the right;M,O,D and S stand for noble Metal,metal Oxides,Dopants and Supports,respectively.Main effects of catalyst compositions on the CO conversion(right),in the absence of H2(a)and in the presence of H2(b).492 D.Farrusseng /Surface Science Reports 63(2008)487–513Fig.5.Blocking structure of 19×19metal binary library (left).Results of 19×19metal binary study (right);synergy was calculated as the difference between the observed TON for the combination minus the sum of the TONs for the individual metals [77].can be measured using an IR focal plane detector [62–65].Other alternatives were developed to overcome selectivity measurement issues such as the use of mass spectrometry (MS)equipped with capillary sampling as shown in [66–70,44,49,71–73].Devices combining IR thermography for rapid hit identification and MS analysis have also been developed [40,74].See also Fig.3.3.3.Design of experiments (DoE)methodology (See also Figs.4and 5)DoE methodology,which involves the simultaneous modifi-cation of variables (usually called factors)and the avoidance of redundant experiments,is widely used in the domain of process engineering.Even though their algorithms are based on simple linear regression,DoE tools can be applied to different objectives that are of particular value in HT experimentation.By means of homogeneous sampling,DoE can be used to efficiently explore a large search space defined by many discrete variables,while guar-anteeing maximal efficiency in terms of the information gleaned from experiments.DoE methodology,when used for screening purposes,quantifies the effect of each individual variable on the targeted properties and identifies the variables relevant to further rounds of screening.On the other hand,when one seeks informa-tion regarding catalytic mechanisms or metal or metal–support synergisms,special DoE design also allows the quantification of in-teractions between variables.Finally,when all relevant variables have been identified,DoE planning is very efficient for the fine op-timization of both catalyst synthesis and process conditions.In this case,the most robust surface responses are generated as empirical models while minimizing the number of experiments.Recent pub-lications illustrate the versatility and power of DoE applied to HT catalyst experimentation.At TU Delft,three sequential DoEs were carried out to find a new one-pot route for the catalytic hydrogenation of acylated cyanohydrins to N -acyl β-amino alcohols [75].Both catalyst formulations and process conditions were varied within the first screening,as shown in Table 2.A selection of 24reactions out of the 320total possible combinations was performed through the use of a D-optimal algorithm.This design is appropriate in view of discarding irrelevant variables for further rounds of optimization.New insights derived from the first design allowed a second one to be carried out using the parameters indicated in Table 3.Variables showing very little impact were discarded.On the other hand,since the nature of the support appeared to bear major importance,a new support,silica,was introduced in the search space.The now considerably reduced parameter space rendered feasible a full-factorial design (i.e.,36combinations),providing accurate information on the main effects of the parameters and especially on the interactions between the parameters.In suchTable 2Definition of the parameter search space for the first screening round.Table 3Definition of the parameter search space for the second screening round.designs,about half the possible number of reactions are typically performed.Finally,a third design was performed in conventional reactors,taking into account other variables such as pressure and temperature.A strategy of sequential hierarchical designs is believed to be better than a single large one,because the information obtained from one design is used to improve the next.If one large DoE design had been chosen,many unnecessary reactions would have been performed.Preliminary designs (typically fewer than 25%of the possible reactions)are sufficient to allow differentiation between significant and insignificant parameters,and are therefore well-suited to reduce the search space in the early stages of the research effort.The search for selective CO oxidation catalysts for H 2purifica-tion applications has also been a context for using a very similar hi-erarchical strategy using a D-optimal design algorithm for primary screening [76].An a priori selection of elements and of combinato-rial rules for mixing them and generating multi-component cata-lysts was performed according to literature data and pre-existing knowledge.Four groups of elements were considered:noble met-als (Pt,Pd,Ru,Rh and Au),oxides (transition metal oxides of Cr,Co,Mn,La,Sm and Mo),dopants (alkali or earth alkali Li,Cs and Ca)and supports (Al 2O 3,CeO 2,ZrO 2,ZnO and C).One of the a priori rules established was that all catalysts were to be composed of one sup-port and two noble metals and,optionally,of one transition metal and one dopant.The weight percentages of noble metal,as well as transition metal and dopant (when present),were fixed at 0.5%,20%and 1%,respectively.The choice to employ two distinct noble metals per catalyst was based on the assumption that alloys may。
